Updates `NoiseModelWithAdditiveBias` to apply per-feature bias sampling (#2760)

# Description

This PR updates the `NoiseModelWithAdditiveBias` to apply per-feature
bias sampling.

Previously, the model sampled a single scalar bias per episode and
applied it uniformly across all feature dimensions (i.e., axis 1 of a
(`num_env`, `feature_dim`) tensor). This PR changes the behavior to
instead sample a separate bias value for each feature dimension, making
the model more suitable for structured inputs such as positions,
velocities, or multi-DOF actions.

### Notes

Structured inputs typically contain semantically distinct components,
like [x, y, z] coordinates, where applying the same bias across all
components introduces unrealistic, fully correlated noise. Independent
per-dimension bias sampling leads to more realistic and robust policy
training, especially for sim-to-real transfer.

I’ve replaced the previous behavior with this new default, as I believe
the original implementation could be misleading and not well-suited for
many practical scenarios. However, if desired, I can make both behaviors
available via a configuration flag (e.g., `per_feature_bias=True`), to
retain backward compatibility.

Fixes #2759.

## Type of change

- Bug fix (non-breaking change which fixes an issue)
- New feature (non-breaking change which adds functionality)

## Checklist

- [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with
`./isaaclab.sh --format`
- [ ] I have made corresponding changes to the documentation
- [x] My changes generate no new warnings
- [ ] I have added tests that prove my fix is effective or that my
feature works
- [x] I have updated the changelog and the corresponding version in the
extension's `config/extension.toml` file
- [x] I have added my name to the `CONTRIBUTORS.md` or my name already
exists there

---------
Co-authored-by: Kelly Guo <kellyg@nvidia.com>

source/isaaclab/config/extension.toml

 [package]
 
 # Note: Semantic Versioning is used: https://semver.org/
-version = "0.40.8"
+version = "0.40.9"
 
 # Description
 title = "Isaac Lab framework for Robot Learning"

source/isaaclab/docs/CHANGELOG.rst

 Changelog
 ---------
 
+0.40.9 (2025-06-25)
+~~~~~~~~~~~~~~~~~~~
+
+Added
+^^^^^
+
+* Added ``sample_bias_per_component`` flag to :class:`~isaaclab.utils.noise.noise_model.NoiseModelWithAdditiveBias` to enable independent per-component bias
+ sampling, which is now the default behavior. If set to False, the previous behavior of sharing the same bias value across all components is retained.
+
+
 0.40.8 (2025-06-18)
 ~~~~~~~~~~~~~~~~~~~
 

source/isaaclab/isaaclab/utils/noise/noise_cfg.py

@@ -103,3 +103,9 @@ class NoiseModelWithAdditiveBiasCfg(NoiseModelCfg):
 
     Based on this configuration, the bias is sampled at every reset of the noise model.
     """
+
+    sample_bias_per_component: bool = True
+    """Whether to sample a separate bias for each data component.
+
+    Defaults to True.
+    """

source/isaaclab/isaaclab/utils/noise/noise_model.py

@@ -154,6 +154,8 @@ class NoiseModelWithAdditiveBias(NoiseModel):
         # store the bias noise configuration
         self._bias_noise_cfg = noise_model_cfg.bias_noise_cfg
         self._bias = torch.zeros((num_envs, 1), device=self._device)
+        self._num_components: int | None = None
+        self._sample_bias_per_component = noise_model_cfg.sample_bias_per_component
 
     def reset(self, env_ids: Sequence[int] | None = None):
         """Reset the noise model.
@@ -179,4 +181,11 @@ class NoiseModelWithAdditiveBias(NoiseModel):
         Returns:
             The data with the noise applied. Shape is the same as the input data.
         """
+        # if sample_bias_per_component, on first apply, expand bias to match last dim of data
+        if self._sample_bias_per_component and self._num_components is None:
+            *_, self._num_components = data.shape
+            # expand bias from (num_envs,1) to (num_envs, num_components)
+            self._bias = self._bias.repeat(1, self._num_components)
+            # now re-sample that expanded bias in-place
+            self.reset()
         return super().__call__(data) + self._bias